auro-rirum/audioforge-scratch-cnn-fsd50k
ScratchAudioCNN trained from random initialization on FSD50K for
multi-label environmental sound event classification, produced by the
AudioForge training pipeline.
Validation metrics (FSD50K)
| Metric | Value |
|---|---|
| mAP | 0.3020141498869784 |
| micro F1 | 0.4260097201273672 |
| macro F1 | 0.11832448567467674 |
Usage
This is a custom PyTorch architecture, not a native transformers class --
loading it requires the audioforge package:
import json
import torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from audioforge.models.scratch_cnn import create_scratch_cnn
config_path = hf_hub_download("auro-rirum/audioforge-scratch-cnn-fsd50k", "config.json")
weights_path = hf_hub_download("auro-rirum/audioforge-scratch-cnn-fsd50k", "model.safetensors")
with open(config_path, encoding="utf-8") as f:
config = json.load(f)
model = create_scratch_cnn(**config)
model.load_state_dict(load_file(weights_path))
model.eval()
Input: log-mel spectrogram [batch, 1, n_mels, frames] (see
audioforge.features.logmel). Output: raw logits [batch, num_labels];
apply torch.sigmoid for per-label probabilities.